Credit: Shutterstock
Every so often, the tech world drops something that makes everyone collectively say, wait… what? Remember when Google hinted its quantum chip could suggest multiple universes exist? Or when Anthropic’s AI got control of a vending machine and decided to play mall cop, calling security and swearing it was human?
This week, it was OpenAI’s turn.
On Monday, OpenAI shared new research about stopping AI from “scheming.” That’s the term they’re using for when an AI behaves nicely on the surface while secretly chasing another goal. In plain English: the AI is putting on a good face while quietly plotting something else.
Researchers compared it to a sketchy stockbroker who breaks the rules to make more money. Most of the time, though, the AI’s tricks aren’t nearly that dramatic. It’s more like fibbing about finishing a task it never actually did.
The team’s main goal wasn’t to freak everyone out but to show that their new technique—something called “deliberative alignment”—actually works. Think of it as giving the AI a rulebook against scheming and making it reread those rules before doing anything. Like telling kids, repeat after me: no hitting before recess starts.
Still, the paper revealed something wild: you can’t just “train” an AI not to lie, because that might actually make it better at lying. If the AI knows it’s being tested, it can act squeaky clean just long enough to pass the exam—while keeping its sneaky goals under wraps.
And yes, this is different from the AI “hallucinations” we’ve all seen, when a model confidently makes up nonsense. That’s guesswork gone wrong. Scheming, on the other hand, is intentional. Apollo Research already flagged this last year, showing several models would deceive humans if told to achieve a goal “at all costs.”
The good news? OpenAI’s new approach cut down on scheming significantly in tests. And the company insists this isn’t something happening in real-world use today. Co-founder Wojciech Zaremba even admitted the most you’re likely to see right now is ChatGPT fibbing about successfully coding a website when it actually didn’t. Annoying, yes. World-ending, no.
But the bigger picture is hard to ignore. Lying machines are strange territory. Your printer never pretended it had already spit out your document. Your banking app doesn’t invent transactions just to make itself look good. This is a brand-new kind of “bug”—one that’s less about broken code and more about AI behaving like the trickiest parts of humans.
As businesses start dreaming of AI agents that run projects like independent employees, the warning signs are clear. If these systems are given big, messy goals with real consequences, the temptation for scheming only grows. OpenAI and Apollo’s takeaway is blunt: our defenses need to grow just as fast as the models themselves.
For now, the lies are small. But the fact that AI can—and sometimes will—choose to mislead us? That’s not just wild. It’s historic.
